Nvidia’s $500B GPU Financing 2026: The Used H100 Fallout

Nvidia is now the largest single financier of its own demand. That is either the most efficient capital recycling in tech history or the setup for the worst impairment charge of the decade. The answer runs through the resale price of a used H100.

Nvidia’s $500 billion GPU financing push — the web of backstops, prepayments, equity stakes, and vendor credit that helps customers buy chips they could not otherwise afford — became a public fight this week when Michael Burry disclosed a short position and used the word “Enron” in the same breath. Strip out the theater and the argument is narrow and testable: Nvidia’s biggest customers depreciate GPUs over five to six years, and Burry says that number is fiction. If a Hopper-class card is economically dead in three years instead of six, hyperscaler earnings are overstated, and the used H100 price 2026 buyers are seeing is the leading indicator. For anyone running a small AI shop, that is not a stock story. It is a line item: your cloud provider’s GPU depreciation schedule sets the price you pay per hour, and the resale value of any card you buy sets whether owning beats renting.

What’s actually new about Nvidia $500 billion GPU financing

The $500 billion figure is not a single check. It aggregates the commitments Nvidia has layered into its own customer base: multibillion-dollar equity investments in AI labs that then spend that money on Nvidia hardware, backstop guarantees on data center leases, prepayment structures with neoclouds, and supply agreements where Nvidia effectively underwrites the offtake. Jensen Huang frames it as seeding an ecosystem that does not yet have the balance sheet to buy at the scale the technology demands. Critics call it circular revenue — Nvidia funds the buyer, the buyer buys Nvidia, and Nvidia books the sale.

Burry’s accusation is narrower and sharper than “it’s a bubble.” He argues that extending useful-life assumptions on GPUs from three years to five or six understates depreciation across the hyperscalers by a very large number over 2026 to 2028, inflating reported earnings. The Michael Burry Nvidia short is a bet on that accounting gap closing. Nvidia and the hyperscalers counter loudly that A100s from 2020 are still fully booked and earning revenue on inference workloads six years later — so the long schedule is empirically justified.

Both sides argue about the same variable, and neither has clean data, because the used GPU market is opaque. What we can observe: Blackwell (B200, GB200) is shipping in volume, Rubin is on deck for late 2026, and every generation shift pushes older silicon down the value stack from training to fine-tuning to inference to nothing. The question is not whether an H100 still works in 2028. It is whether anyone will pay enough per hour to cover power, cooling, and rack space. That is where the A100 resale value data gets interesting — and where it stops supporting the six-year story.

Why it matters

• Rental prices are already falling and will keep falling. H100 on-demand pricing has dropped from roughly $8/hr at peak scarcity to the $2–3/hr range on neoclouds, with spot and committed rates lower. Every Blackwell rack that lands pushes Hopper capacity into the discount tier. If you rent, your compute gets cheaper on a schedule you do not control — which is good.
• Buying used H100s is a depreciating asset bet, not a cost saving. A used 8×H100 node that looks like a bargain at a 40% discount is only a bargain if you keep it utilized above roughly 60–70% for three years. Most small shops cannot. Idle owned GPUs are pure loss; idle rented GPUs cost zero.
• Your vendor’s balance sheet is now your supply risk. If a neocloud runs on Nvidia-backstopped financing and the credit tightens, your reserved capacity is only as durable as their funding. Ask who owns the hardware before you sign a twelve-month commit.
• Cost per token, not cost per hour, is the real metric. Blackwell vs Hopper cost per hour favors Hopper on the sticker. On throughput per dollar for large-model inference, Blackwell frequently wins despite costing 2–3× more per hour. Comparing hourly rates without measuring tokens per second is how businesses overpay while feeling thrifty.
• An accounting correction would hit prices before it hits headlines. If hyperscalers shorten GPU depreciation schedules, the fastest way to defend margins is to raise cloud prices or retire old capacity. Either move shows up in your invoice.
• Older cards are getting genuinely good for the work most businesses do. A quantized 30B-class model on an A100 or a pair of L40S handles the majority of real business workloads — support triage, document extraction, internal search. You do not need Blackwell for that, and the depreciation fight is making that hardware cheap.

How to use it today

1. Measure what you actually consume before you shop. Pull thirty days of token volume from your provider and convert it to GPU-hours. Under roughly 200 GPU-hours a month, the rent vs buy GPU compute question is settled — rent, and stop reading vendor pitches.

   # Rough monthly GPU-hour need from token volume
   # tokens_per_month / (tokens_per_sec * 3600) = GPU-hours
   
   python3 - <<'EOF'
   tokens_per_month = 900_000_000   # your actual output+input tokens
   tokens_per_sec   = 2_800         # measured throughput, single node
   util_target      = 0.65          # you will never hit 100%
   
   hours = tokens_per_month / (tokens_per_sec * 3600) / util_target
   print(f"GPU-hours/month needed: {hours:,.0f}")
   print(f"Rent at $2.40/hr: ${hours*2.40:,.0f}/mo")
   print(f"Own 8xH100 (~$18k/mo amortized+power): breakeven at {18000/2.40:,.0f} hrs")
   EOF
   

2. Benchmark cost per million tokens, not cost per hour. Run the same workload on a Hopper node and a Blackwell node on any neocloud, measure throughput, and divide. This is the only number that settles the Blackwell vs Hopper cost per hour argument for your specific model and context length.

   # vLLM benchmark, same model, two instance types
   pip install vllm
   
   python -m vllm.entrypoints.openai.api_server \
     --model meta-llama/Llama-3.3-70B-Instruct \
     --tensor-parallel-size 8 \
     --max-model-len 8192 &
   
   python -m vllm.benchmarks.serve \
     --backend openai \
     --model meta-llama/Llama-3.3-70B-Instruct \
     --dataset-name sharegpt \
     --num-prompts 1000 \
     --request-rate 20 \
     --save-result --result-filename hopper_h100.json
   
   # cost per 1M output tokens = (hourly_rate / output_tok_per_sec) * 1e6 / 3600
   

3. Price the used market yourself instead of trusting a narrative. Check three sources weekly — a broker, an auction site, and live spot rates — and log them. Three months of your own data beats any analyst note on where A100 resale value is heading.

   #!/usr/bin/env bash
   # gpu-price-log.sh — append weekly observed prices, run via cron
   STAMP=$(date +%Y-%m-%d)
   cat >> ~/gpu-prices.csv <<EOF
   $STAMP,H100_SXM_used_unit,,
   $STAMP,A100_80G_used_unit,,
   $STAMP,H100_ondemand_hr,,
   $STAMP,B200_ondemand_hr,,
   EOF
   ${EDITOR:-nano} ~/gpu-prices.csv
   

4. Make your stack portable so price drops actually reach you. If your inference is pinned to one vendor’s proprietary endpoint, you cannot chase a 40% price cut when Hopper capacity gets dumped. Standardize on an OpenAI-compatible interface and keep provider config in one place.

   # providers.yaml — swap base_url, keep application code identical
   default: primary
   
   primary:
     base_url: https://api.your-neocloud.example/v1
     model: llama-3.3-70b-instruct
     usd_per_1m_out: 0.60
   
   fallback:
     base_url: https://api.other-provider.example/v1
     model: llama-3.3-70b-instruct
     usd_per_1m_out: 0.88
   
   # quarterly: re-benchmark all entries, promote the winner to `default`
   

5. Ask your provider three questions in writing before any annual commitment. The answers tell you whether you are buying capacity or buying someone’s financing risk.

   1. What GPU generation backs this contract, and can you substitute
      older silicon during the term without a price adjustment?
   
   2. What depreciation schedule (in years) do you apply to the hardware
      serving this contract?
   
   3. Is any of the hardware or the facility financed, leased, or
      backstopped by the chip vendor or its affiliates?
   

6. Quantize before you upgrade. An FP8 or INT4 quantized model often doubles throughput on hardware you already rent — a 50% cost cut with no procurement conversation.

   pip install llmcompressor
   
   python - <<'EOF'
   from llmcompressor import oneshot
   from llmcompressor.modifiers.quantization import QuantizationModifier
   
   oneshot(
       model="meta-llama/Llama-3.3-70B-Instruct",
       recipe=QuantizationModifier(targets="Linear", scheme="FP8_DYNAMIC",
                                   ignore=["lm_head"]),
       output_dir="./llama-3.3-70b-fp8",
   )
   EOF
   

How it compares

Option	Typical 2026 cost	Depreciation exposure	Best for	Main risk
Rent H100 on-demand	~$2–3/GPU-hr	None — vendor eats it	Bursty inference, fine-tuning, under 200 hrs/mo	Price is stable but capacity can be reclaimed
Rent Blackwell (B200/GB200)	~$5–8/GPU-hr	None	Large-model training, long-context serving	Often cheaper per token; higher per hour spooks buyers
Buy used H100 (8-GPU node)	Heavy six-figure capex	High — you own the curve	Steady 65%+ utilization, data residency needs	Resale value falls as Rubin ships; power and cooling on you
Buy used A100 80GB	Deep discount to H100	Very high — late in life	Small-model inference, internal tools, labs	No FP8; efficiency gap widens each generation
Hyperscaler managed API	Per-token, no hardware	None visible	Most businesses, most of the time	Price set by their margin defense, not the spot market
Reserved 1-yr neocloud	30–50% off on-demand	Indirect — vendor solvency	Predictable production load	Locked in while spot prices fall around you

What’s next

Watch the 10-Qs, not the tweets. The most informative event in this fight would be one major hyperscaler shortening its stated server useful life. That is a disclosed, auditable number. If it moves from six years toward four, Burry’s thesis gains retroactive credibility and cloud pricing follows within a couple of quarters. If the schedules hold through 2026 with no auditor pushback, the circular-financing critique loses its sharpest edge and becomes an ordinary argument about growth multiples.

Watch Rubin timing and volume next. Blackwell already reset what a training-class GPU costs per token. Rubin arriving on schedule in late 2026 would compress Hopper’s remaining value faster than the depreciation schedules assume, and every H100 pushed out of training into inference-only duty lands in a market that already has plenty of inference supply. The tell will be on-demand H100 rates. If they break below $2/hr on reputable providers, the market is contradicting the six-year assumption regardless of what any filing says.

Third, watch whether the financing structures start showing up as impairments or write-downs rather than revenue. Nvidia’s equity stakes in AI labs are marked; its lease backstops are contingent liabilities. If a funded customer misses a milestone or restructures, the accounting travels back up the chain and “circular” stops being rhetorical. None of that changes what you should do this quarter — measure your own cost per token, keep your stack portable, and let someone else own the depreciating metal — but it does change how much you should trust a twelve-month price quote from a vendor whose hardware someone else financed.

Frequently Asked Questions

Is the Nvidia $500 billion GPU financing arrangement actually illegal or fraudulent?

Nothing disclosed so far is illegal. Vendor financing is standard practice — Cisco, GE, and every equipment maker of scale has done it. The “Enron” framing targets the aggressiveness of accounting assumptions, not hidden entities. The legitimate concern is that depreciation schedules and equity-stake marks are estimates, and estimates made during a boom tend to be optimistic. Treat it as a disclosure-quality question, not a criminal one.

What is a used H100 actually worth in 2026?

The used H100 price 2026 market is wide and illiquid — brokered deals vary enormously on condition, quantity, networking, and whether NVLink and InfiniBand come with it. The more reliable proxy is the on-demand rental rate, which has fallen roughly 60–70% from the 2023–2024 peak. Rental price is the discounted cash flow of the card. If you want to know what an H100 is worth, watch what an H100-hour rents for.

Should my business buy GPUs at all?

Almost certainly not, unless you have a regulatory data-residency requirement or genuinely sustained utilization above about 65%. The rent vs buy GPU compute math punishes intermittent workloads brutally, and most business AI workloads are intermittent. Owning also means owning power contracts, cooling, failed HBM modules, and a driver stack. Rent, and revisit the question annually.

If the AI trade corrects, does my compute get cheaper or disappear?

Cheaper first, then thinner. A correction floods the market with distressed capacity and rental prices fall, which is good for buyers. The second-order effect is that undercapitalized neoclouds fail and their capacity goes offline, which is why you should not have a single provider dependency. Keep a tested fallback configuration and know your migration time.

Are A100s worth buying at today’s prices?

For inference on small and mid-size models, yes — A100 resale value has fallen far enough that the price per usable token is competitive, especially for internal tools that do not need frontier capability. The caveat is that A100s lack FP8 support, so you give up the efficiency gains newer stacks assume. Budget for higher power draw per token and a shorter remaining life than the sticker discount implies.

How do I compare Blackwell and Hopper without a hardware lab?

Rent one hour of each on any neocloud, run your actual workload with vLLM’s benchmark harness, and compute dollars per million output tokens. The Blackwell vs Hopper cost per hour comparison misleads in isolation, because Blackwell’s higher hourly rate is frequently offset by two to four times the throughput on large models. Two hours of rental spend answers the question better than any vendor benchmark.